Downloaded from rnajournal.cshlp.org on September 27, 2021 - Published by Cold Spring Harbor Laboratory Press BIOINFORMATICS A common class of transcripts with 5′-intron depletion, distinct early coding sequence features, and N1-methyladenosine modification CAN CENIK,1,2 HON NIAN CHUA,3,4,5 GURAMRIT SINGH,2,6,7,8 ABDALLA AKEF,9 MICHAEL P. SNYDER,1 ALEXANDER F. PALAZZO,9 MELISSA J. MOORE,2,7,8 and FREDERICK P. ROTH3,4,10,11 1Department of Genetics, Stanford University School of Medicine, Stanford, California 94305, USA 2Department of Biochemistry and Molecular Pharmacology, University of Massachusetts Medical School, Worcester, Massachusetts 01605, USA 3Donnelly Centre, Department of Molecular Genetics, and Department of Computer Science, University of Toronto, Toronto M5S 3E1, Ontario, Canada 4Lunenfeld-Tanenbaum Research Institute, Mt. Sinai Hospital, Toronto M5G 1X5, Ontario, Canada 5DataRobot, Inc., Boston, Massachusetts 02109, USA 6Department of Molecular Genetics, Center for RNA Biology, The Ohio State University, Columbus, Ohio 43210, USA 7Howard Hughes Medical Institute, University of Massachusetts Medical School, Worcester, Massachusetts 01605, USA 8RNA Therapeutics Institute, University of Massachusetts Medical School, Worcester, Massachusetts 01605, USA 9Department of Biochemistry, University of Toronto, Toronto, Ontario M5S 1A8, Canada 10Center for Cancer Systems Biology (CCSB), Dana-Farber Cancer Institute, Boston 02215, Massachusetts, USA 11The Canadian Institute for Advanced Research, Toronto M5G 1Z8, Ontario, Canada ABSTRACT Introns are found in 5′ untranslated regions (5′UTRs) for 35% of all human transcripts. These 5′UTR introns are not randomly distributed: Genes that encode secreted, membrane-bound and mitochondrial proteins are less likely to have them. Curiously, transcripts lacking 5′UTR introns tend to harbor specific RNA sequence elements in their early coding regions. To model and understand the connection between coding-region sequence and 5′UTR intron status, we developed a classifier that can predict 5′UTR intron status with >80% accuracy using only sequence features in the early coding region. Thus, the classifier identifies transcripts with 5′ proximal-intron-minus-like-coding regions (“5IM” transcripts). Unexpectedly, we found that the early coding sequence features defining 5IM transcripts are widespread, appearing in 21% of all human RefSeq transcripts. The 5IM class of transcripts is enriched for non-AUG start codons, more extensive secondary structure both preceding the start codon and near the 5′ cap, greater dependence on eIF4E for translation, and association with ER-proximal ribosomes. 5IM transcripts are bound by the exon junction complex (EJC) at noncanonical 5′ proximal positions. Finally, N1-methyladenosines are specifically enriched in the early coding regions of 5IM transcripts. Taken together, our analyses point to the existence of a distinct 5IM class comprising ∼20% of human transcripts. This class is defined by depletion of 5′ proximal introns, presence of specific RNA sequence features associated with low translation efficiency, N1-methyladenosines in the early coding region, and enrichment for noncanonical binding by the EJC. Keywords: 5′-UTR introns; random forest; N1-methyladenosine; exon junction complex INTRODUCTION scripts that encode ER- and mitochondria-targeted proteins, 5UI depletion is associated with the presence of specific RNA Approximately 35% of all human transcripts harbor introns ′ ′ sequence properties (Palazzo et al. 2007, 2013; Cenik et al. in their 5 untranslated regions (5 UTRs) (Hong et al. 2006; ′ 2011). Specifically, nuclear export of an otherwise inefficient- Cenik et al. 2010). Among genes with 5 UTR introns ly exported microinjected mRNA or cDNA transcript can be (5UIs), those annotated as “regulatory” are significantly over- promoted by an ER-targeting signal sequence-containing re- represented, while there is an underrepresentation of genes gion (SSCRs) or mitochondrial signal sequence coding re- encoding proteins that are targeted to either the endoplasmic ′ gion (MSCRs) from a gene lacking 5 UTR introns (Cenik reticulum (ER) or mitochondria (Cenik et al. 2011). For tran- et al. 2011; Lee et al. 2015). However, more recent studies suggest that many SSCRs have little impact on nuclear export Corresponding authors: [email protected], Fritz.Roth@ utoronto.ca © 2017 Cenik et al. This article, published in RNA, is available under a Article is online at http://www.rnajournal.org/cgi/doi/10.1261/rna.059105. Creative Commons License (Attribution 4.0 International), as described at 116. Freely available online through the RNA Open Access option. http://creativecommons.org/licenses/by/4.0/. 270 RNA 23:270–283; Published by Cold Spring Harbor Laboratory Press for the RNA Society Downloaded from rnajournal.cshlp.org on September 27, 2021 - Published by Cold Spring Harbor Laboratory Press Defining a new mRNA class with distinct features for RNAs transcribed in vivo (Lee et al. 2015), but rather en- the absence of a 5UI, we sought to model this relationship. hance translation in a RanBP2-dependent manner Specifically, we used a random forest classifier (Breiman (Mahadevan et al. 2013). 2001) to learn the relationship between 5UI absence and a Among SSCR- and MSCR-containing transcripts (referred collection of 36 different sequence features extracted from to hereafter as SSCR and MSCR transcripts), ∼75% lack the first 99 nt of all human coding regions (CDS) (Fig. 1A– − 5′UTR introns (“5UI ” transcripts) and ∼25% have them C; Supplemental Table S1; Materials and Methods). We then (“5UI+” transcripts). These two groups have markedly differ- used all transcripts known to contain an SSCR (a total of ent sequence compositions at the 5′ ends of their coding se- 3743 transcripts clusters; Materials and Methods), regardless − quences. 5UI transcripts tend to have lower adenine content of 5UI status, as our training set. This training constraint en- (Palazzo et al. 2007) and use codons with fewer uracils and sured that all input nucleotide sequences were subject to sim- adenines than 5UI+ transcripts (Cenik et al. 2011). Their sig- ilar functional constraints at the protein level. Thus, we nal sequences also contain leucine and arginine more often sought to identify sequence features that differ between − than the biochemically similar amino acids isoleucine and ly- 5UI and 5UI+ transcripts at the RNA level. sine, respectively. Leucine and arginine codons contain fewer Our classifier assigns to each transcript a “5′UTR-intron- adenine and thymine nucleotides, consistent with adenine minus-predictor” (5IMP) score between 0 and 10, where and thymine depletion. This depletion is also associated higher scores correspond to a higher likelihood of being − with the presence of a specific GC-rich RNA motif in the ear- 5UI (Fig. 1C). Interestingly, preliminary ranking of the − − ly coding region of 5UI transcripts (Cenik et al. 2011). 5UI transcripts by 5IMP score revealed a relationship be- Despite some knowledge as to their early coding region tween the position of the first intron in the coding region − − features, key questions about this class of 5UI transcripts and the 5IMP score. 5UI transcripts for which the first in- have remained unanswered: Do the above sequence features tron was more than 85 nt downstream from the start codon extend beyond SSCR- and MSCR-containing transcripts to had the highest 5IMP scores. Furthermore, the closer the first − − other 5UI genes? Do 5UI transcripts having these features intron was to the start codon, the lower the 5IMP score share common functional or regulatory features? What bind- (Fig. 1D). We explored this relationship further by training ing factor(s) recognize these RNA elements? A more com- classifiers that increasingly excluded from the training set − plete model of the relationship of early coding features and 5UI transcripts according to the distance of the first intron − 5UI status would begin to address these questions. from the 5′ end of the coding region. This revealed that Here, to better understand the relationship between early classifier performance, as measured by the area under the coding region features and 5UI status, we undertook an precision recall curve (AUPRC), increased as a function of integrative machine learning approach. We reasoned that a the distance from start codon to first intron distance − machine learning classifier which could identify 5UI tran- (Materials and Methods, Fig. 1E). Thus, the RNA sequence − scripts solely from early coding sequence would potentially features we identified as being predictive of 5UI transcripts provide two types of insight. First, it could systematically are more accurately described as being predictors of tran- − identify predictive features. Second, the subset of 5UI tran- scripts without 5′-proximal introns. scripts that could be identified by the classifier might then To minimize the impact of transcripts that may “behave” represent a functionally distinct transcript class. Having as though they were 5UI+ due to an intron early in the − developed such a classifier, we found that it identified coding region, we eliminated 5UI SSCR transcripts with ∼21% of all human transcripts as harboring coding regions a first intron <90 nt downstream from the start codon − characteristic of 5UI transcripts. While many of these tran- (Materials and Methods) and generated a new classifier. scripts encode ER- and mitochondrial-targeted proteins, Discriminative motif